Research

Foundation models & publications

Multimodal foundation models for documents, images and interfaces, plus recent work on evaluating frontier language models.

Areas#

Vision language

Models that see and read at once. LaTr for scene text VQA, MixGen for multimodal data augmentation and YORO for efficient end to end visual grounding.

Computer Use GUI Agents

Agents that operate real interfaces. R-VLM for precise GUI grounding at ACL 2025 and compressed history states that make web automation far faster and cheaper.

Foundation models

Pretrained backbones for documents and the visual world. DocFormer and DocFormerv2 for end to end document understanding, with rich supervision and self supervised pretraining underneath.

Mentees & Collaborators#

I am fortunate and proud to have worked with incredible researchers. I have learnt from them much more than what I could teach back.

Yash Patel·learnt image compression·scientist at Meta
Poojan Oza·research into multimodal OCR training·scientist at Amazon Go
Chih-Hui (John) Ho·vision language research, ECCV·scientist at SAP
Xiaoshuai Hao·vision language research·researcher at Xiaomi EV
Ali Furkan Biten·vision language research, CVPR·at Computer Vision Center, Barcelona
Yichu Zhou·document understanding with LLMs·scientist at Pinterest
Chenge Li·vision language research·senior scientist at Amazon
Yuan Gao·computer use vision language pretraining, CVPR·researcher at xAI
Park Joonhyung·precise GUI grounding, ACL·postdoc at KAIST
Varun Nagaraj Rao·retrieval augmented vision language learning·PhD at Princeton
Tian Li·accelerated inference in transformer models·scientist at Amazon
Ying Nian Zhao·multimodal transformer distillation·scientist at Amazon
Xiyue Zhu·improving computer use automation with compressed history states·PhD at UIUC
Shrinidhi Kumbhar·accelerating inference with diffusion LLMs, CVPR·PhD at Arizona State
Sungnyun Kim·distillation on document understanding·postdoc at KAIST
Simone Bombari·differential privacy research·PhD at ISTA
Ofir Abramovich·multimodal research·PhD at Reichman University
Tianyang Zhao·multimodal distillation research·scientist at Amazon Go

Selected Publications#

Preprint 2026 RewardHack-Bench: A Benchmark for Reward Hacking Detection in vLLMs with Price, Benton and Hubinger

ICLR 2026 ImpossibleBench: Measuring LLMs' Propensity of Exploiting Test Cases

arXiv 2025 Turbocharging Web Automation: The Impact of Compressed History States with Zhu, Tang and Liao

ACL Findings 2025 R-VLM: Region-Aware Vision Language Model for Precise GUI Grounding

AAAI 2024 DocFormerv2: Local Features for Document Understanding 100+ citations

CVPR 2024 Enhancing Vision-Language Pre-training with Rich Supervisions

WACV 2023 MixGen: A New Multi-Modal Data Augmentation 175+ citations

CVPR 2022 LaTr: Layout-Aware Transformer for Scene-Text VQA 140+ citations

ECCV 2022 YORO: Lightweight End to End Visual Grounding

ICCV 2021 DocFormer: End-to-End Transformer for Document Understanding 580+ citations

WACV 2021 Saliency Driven Perceptual Image Compression 70+ citations

Full list on Google Scholar.